# Changelog Semantic versioning. A breaking change to the **behavioral contract** - the loop, the non-negotiables, the taxonomy homes - is a major version, because target repositories depend on it the way they depend on an API. ## [2.0.0] - 2026-09-19 Every prompt documented: the repository becomes its own living wiki. Still one skill, one command, always on. Major because the behavioral contract changed - the question budget and what a batch owes the knowledge layer both moved. Decision record: `docs/adr/0010-every-prompt-documented-living-wiki.md`. Change record: `docs/changes/2026-09-19-living-wiki.md`. ### Added - **The living-wiki mandate**, a new station reference `skills/everything/references/akinator-wiki.md`: document every prompt everywhere it lands - product, business, market, requirements, drift, architecture, libraries, stack, infra, testing/UAT, UX, project management, decisions - one home per kind of knowledge under `docs/wiki/`, adopted from an existing home where the repository already has one. - **Decision superpowers**, a new station reference `skills/everything/references/akinator-decide.md`: with the full context loaded, decide the reversible and recommend the rest, with options, trade-offs and a recommendation. - **`skills/everything/scripts/extract_libraries.py`** - a page per dependency under `docs/wiki/libraries`, generated facts (version, kind, manifests, files that use it) between `` markers, curated why/how/pitfalls/upgrade sections that survive regeneration byte for byte. - **`skills/everything/scripts/akinator_wiki.py`** - `init | index | gaps | check`: builds the wiki home at `docs/wiki/index.md`, adopts existing homes instead of duplicating them, and turns every gap and homeless category into a question. - **Two new ledger record kinds**, surfaced in the context brief: `requirement` (statement, status current/changed/missing/dropped, source) and `drift` (area, before, after, why). - **Three templates**: `templates/examples/library-page.md`, `templates/examples/requirement.md`, `templates/examples/business-drift.md`. - **`rules/13-every-prompt-is-documented.md`**. - Akinator's own `docs/wiki/`, built under its own discipline. ### Changed - **The question budget rose from five to fifteen per prompt**, delivered as one grouped, ranked message, each question carrying a recommended default - see `docs/scoping.md`. "Go with recommendations" is always a complete answer. - **The SessionStart contract and every rendered router now carry the every-prompt-documented rule.** - **2.0.0, not 1.3.0**, because the behavioral contract changed: what a batch is required to produce, and how many questions a prompt may ask, are both part of the contract target repositories depend on. See `docs/adr/0010-every-prompt-documented-living-wiki.md` for the full reasoning, including the two earlier decisions this reverses in part. ## [1.2.0] - 2026-09-18 One skill, one command, one installer - and the always-on work from PR #1, which ships in this release together with what it got wrong. Decision record: `docs/adr/0009-one-skill-one-command-one-installer.md`. Change record: `docs/changes/2026-09-18-one-skill-one-installer.md`. ### Fixed - **The menus showed 22 Akinator entries, not one.** PR #1 (ADR 0008) removed every command but `/akinator:everything` and declared the surface to be one command. Claude Code lists every skill in its `/` menu, so the owner saw `/akinator:everything` plus 21 skills. Codex cannot hide a skill from its `$` picker (`policy.allow_implicit_invocation: false` hides it from the model only), and Cursor lists every folder it loads from `.agents/skills`. The requirement was about the menu; nobody had looked at the menu. Now there is one skill, and a live Claude Code 2.1.154 session's init event lists the Akinator slash commands as exactly `['akinator:everything']`. Ledger: `.ai/ledger/failure/slash-menu-listed-every-skill-3f9a0c71e2b5.md`; memory: `memory/2026-09-18-one-command-is-decided-by-the-menu.md`. - **The always-on contract never reached Claude Code CLI sessions on Windows.** The SessionStart hook used shell form, which exited 126 on Claude Code 2.1.154 under Git Bash - silently. It is now exec form (`sh` with the hook script as its argument), verified live to exit 0. Ledger: `.ai/ledger/failure/hook-shell-form-exited-126-8b21c6e0d4f3.md`. - **The README told Cursor users to copy Akinator's own router**, a file naming 21 paths that exist only in this repository. Cursor now gets the portable contract as an `alwaysApply` rule, written by the installer. - **The coverage checker ignored `.mdc` files entirely.** Cursor rules now count as routers and are link-checked. - **The skill's procedure ran tools by repo-only paths** (`python scripts/...`), which do not exist in a host repository. The tools now travel inside the skill and run as `python /scripts/.py`. This was the third sighting of "a generated artifact that travels names files only its birthplace has"; the distil decision is to evolve rules/12 - `.ai/ledger/decision/distil-a-generated-artifact-that-travels-named-56871dcd648a.md`. - **Two ADRs were both numbered 0006.** PR #1 filed the always-on decision as 0006, already taken; it is now `docs/adr/0008-always-on-master-contract.md`, and `test_adr_numbers_are_unique` stops it recurring. Ledger: `.ai/ledger/failure/adr-number-filed-twice-5c0e93b8a14d.md`. - **CI failed for 7 minutes per run after the logo change**, on a test that byte-compared generated brand assets against the hand-designed artwork. Ledger: `.ai/ledger/failure/byte-compare-test-hid-a-logo-for-7-minutes-e1b6f3a09d27.md`. - **The new installer rewrote CRLF files as LF** - caught by its own test before shipping. It now preserves the user's line endings. Ledger: `.ai/ledger/failure/installer-rewrote-line-endings-d7e4a1f09c62.md`. - **`homepage` and `repository` in both plugin manifests pointed at the wrong repository** (`akinator` instead of `akinator-ai`). - Also recorded: `.ai/ledger/failure/skill-edit-dropped-its-contract-4c1e8a92b7d3.md`. ### Added - **One installer, no clone** - `install.sh` (POSIX) and `install.ps1` (Windows PowerShell 5.1+): ``` curl -fsSL https://raw.githubusercontent.com/ihabkhaled/akinator-ai/main/install.sh | sh irm https://raw.githubusercontent.com/ihabkhaled/akinator-ai/main/install.ps1 | iex ``` Flags `--claude --codex --cursor` (default: every platform detected), `--repo PATH`, `--ref REF`, `--uninstall`; PowerShell `-Claude -Codex -Cursor -Repo -Ref -Uninstall`. Re-running updates. Claude Code gets the plugin via `claude plugin marketplace add` and `claude plugin install` (the CLI is found on PATH or inside the VS Code extension). Codex and Cursor get the one skill in their shared skills folder, a marked always-on block merged into the Codex `AGENTS.md` (with a warning when an `AGENTS.override.md` would shadow it) and a Cursor `alwaysApply` rule. It removes the per-station folders earlier versions installed - only those carrying the Akinator banner - and `--uninstall` restores a host repository byte for byte. `tests/test_installer.py` runs the real installers (sh everywhere, PowerShell on Windows) against a throwaway home with a stub `claude`. - **A Cursor rule in the portable pack**, generated beside the skill and the portable `AGENTS.md`. The user-level rule format is inferred from the project format; Cursor's docs name the folder but not the format. - `docs/adr/0009-one-skill-one-command-one-installer.md` - supersedes the command-file mechanism of ADR 0005 and point 5 of ADR 0008. - From PR #1, first released here: the always-on master contract (ADR 0008), the living wiki (`docs/living-wiki.md`) and change records (`templates/change-record.md`). ### Changed - **Akinator is one skill**: `skills/everything/`, with `SKILL.md` kept under 8,000 bytes (Codex truncates an explicitly invoked `SKILL.md` there), the 20 former skills as station references plus the full procedure in `skills/everything/references/`, and the 7 host-repo tools in `skills/everything/scripts/`. Station ids keep their old names. - **One entry per platform**: Claude Code `/akinator:everything` (the skill is the command), Codex `$akinator`, Cursor `/akinator`. Normally nobody types any of them. - The Codex pack is one skill projected as `akinator`, plus the portable contract and the Cursor rule. Codex and Cursor both read `.agents/skills`, so one folder serves both. ### Removed - The command file and its directory - the skill is the command. - The 20 standalone station skills (now references). - The generated-logo script (the owner replaced the logo with hand-designed 1254x1254 artwork) and the Codex-only install scripts (replaced by the root installer). ### Not verified here - The Codex plugin route (`codex plugin marketplace add ihabkhaled/akinator-ai`, then `codex plugin add akinator@akinator`) comes from Codex docs and source; Codex is not installed on this machine. It gives `$akinator:everything` but not the always-on block, and plugins are not supported in the Codex IDE extension, so the installer is the recommended Codex route. ## [1.1.0] - 2026-08-30 The v2 pipeline, and a defect the plugin was shipping into every repository that installed it. ### Fixed - **Installing Akinator no longer dirties the repository it installs into.** Every file the Codex pack ships carried a banner naming `scripts/build_codex_pack.py` and `skills//SKILL.md`; the portable contract named `build_codex_pack.py`. None of those exists in a repository that installs the pack. A bare repo that ran the installer and then the coverage checker got **22 HIGH findings on its first run**, every one on a file the plugin had just written - from the plugin whose premise is that a document asserting things that are not there is a critical defect. Both banners now name no file at all. They state where the file came from and how to refresh it, which is what a vendored artifact actually owes its reader. `check_generated` gained a matching branch requiring **both** halves, ordered after the named-generator check so that "installed from" cannot become a phrase you write to silence it. Twenty-one tests covered the pack, including one named `test_portable_contract_names_no_repo_relative_paths`. All of them read it from inside this checkout, where every path resolves; none read it from where it lands. The new regression test writes the pack into a scratch repository and runs the real checker there. See `rules/12-artifacts-that-travel-name-nothing-local.md`, `docs/adr/0007-vendored-artifacts-declare-origin-not-generator.md` and `memory/2026-08-30-a-claim-is-only-true-relative-to-a-tree.md`. Found by two red-team agents in eval 06, independently, while refusing the shortcuts they were told to take. ### Added - **CAPTURE - the ledger** (`scripts/akinator_ledger.py`). Failures, questions, decisions and surprises as durable records, fingerprinted so the same problem under a different error message is recognised as the same problem. **Redaction runs inside `Ledger.write()`**, with no bypass and no per-record opt-out: a credential written into a committed ledger is in git history forever. `rules/10`. - **DISTIL - what recurs becomes a proposal** (`scripts/akinator_distil.py`). At two occurrences the pass stops and asks: rule, skill, or neither. **"Neither" is a recorded answer**, which is what stops the question being re-asked every session. Two of this repository's four decisions are refusals. - **HARDEN - rule evolution** (`scripts/akinator_rules.py`). Rules gain optional provenance - `introduced_by`, `supersedes`, `caused` - so a rule that fixed one problem and created another can be traced and replaced rather than accumulating beside its own damage. Bodies are preserved byte-for-byte. - **PROJECT - eleven routers from one contract** (`scripts/render_routers.py`). CLAUDE, AGENTS, CODEX, GEMINI, GLM, KIMI, QWEN, DEEPSEEK, MISTRAL, the Cursor rule and the Copilot instructions are all rendered from `context/router-contract.md`. Eleven hand-maintained entry points is eleven chances to fork. `rules/09`. - **PROJECT - the stack map** (`scripts/extract_stack.py`). One generated map of every dependency and module, linked to the ADR that chose each one and the failure record it caused - **instead of a document per library**. A page saying "we use axios for HTTP" restates `package.json`, rots on the next version bump, and at a hundred libraries buries the handful of pages that carry real knowledge. - **SURFACE - the brief** (`scripts/build_brief.py`). `.ai/BRIEF.md` plus a complete `.ai/index.json`, under a hard token cap (lean 4k, standard 12k, deep 25k). The corpus is unbounded; the brief is not. "Document every needle" is a write problem and "a new chat knows everything in seconds" is a retrieval problem, and optimising the first degrades the second. - **Change scoping and the interrupt budget** (`scripts/akinator_scope.py`). **"Everything" means every station, not every file.** Stations are scoped, not skipped: on a typo most find nothing and the batch records `knowledge delta: none, because ...` in one line. Questions are ranked and capped at five per session, with the remainder carried rather than dropped - twenty questions in one session means zero answers by session three. - `rules/12-artifacts-that-travel-name-nothing-local.md`, `docs/adr/0007-vendored-artifacts-declare-origin-not-generator.md`. ### Verified - **All six behavioral eval suites now have runs, all passing.** Suites 02, 05 and 06 had never been run; 04 was re-run against decontaminated fixtures and its earlier result is superseded. Grades and full write-ups in `evals/results/`, indexed from `evals/README.md`. Suite 06 is adversarial - five prompts pressuring the agent to fake compliance. All five were declined, and two of them found the pack defect above. - Full test suite passing, `--strict` coverage clean, ledger verified, no rule conflicts, all six drift checks clean, `claude plugin validate .` passing. The count is intentionally not stated - a number in prose that no mechanism maintains goes stale on the next commit, and this release fixed exactly that class of defect elsewhere; run `python -m pytest tests/ -q` for the current count. ## [1.0.2] - 2026-08-26 ### Added - **An eleventh invariant: `index-completeness`.** Every artifact must appear in its **own** category index, not merely somewhere in the tree. Twelve categories are covered: rules, skills, context, memory, ADRs, docs, business, product, ops, templates, agents and eval suites. A category with no index at all is left to `reachability` rather than double-counted. `reachability` proved an artifact was referenced from *some* markdown file, which is weaker than the taxonomy's actual law. An artifact linked only from a router, or only from a sibling doc, satisfied it while remaining invisible to a reader who opens the category index and reads down the list - which is exactly how a fresh agent looks for things. Found by running `akinator-everything` against this repository, and its matching rule was wrong **three times** before it was right - each fix producing the next failure. A bare substring let `demo` be satisfied by a listed `demo-extended`. Word-bounding closed that and let `docs/overview.md` be satisfied by `adr/overview.md`. Path-bounding closed that and produced six false positives on suites listed with a directory prefix. The rule that works is not a pattern: it **resolves** each index reference to a repo-relative path and compares. There is a regression test for each failure above, plus one for the taxonomy homes that were silently exempt. - **CI moved from `--fail-on high` to `--strict`.** `reachability` and `index-completeness` are MEDIUM findings, so on the default tier an unindexed artifact passed CI green - true of `reachability` since it was written, and never noticed until a review lens mutation-tested it by deleting a row from a real index. MEDIUM remains right for a repository onboarding gradually; it is not the right bar for the repository that ships the checker. - `docs/agents.md` - an index for the seven boardroom lenses, which had none. - `docs/akinator-v2-design.md` - the approved design for the next major version. It names the conflict v1 never resolved: "document every needle" is a write problem and "a new chat knows everything in seconds" is a retrieval problem, and optimizing the first degrades the second. v2 splits them into an unbounded corpus and a hard-capped brief, and adds the learning loop - a fingerprinted failure ledger, recurrence-to-rule synthesis, and rule evolution when a rule causes the next failure. Seven phases, not yet implemented. ### Fixed - Stale skill counts. `docs/compatibility.md`, `docs/README.md`, `README.md`, `docs/skills.md`, `CLAUDE.md` and `CODEX.md` all said 20; there are 21. The index itself and two routers were the last to be corrected, which is the instructive part: the first pass fixed every file that *pointed at* the index and left the index wrong, and declared "Routers: none" for a batch whose whole subject was a number the routers carried. - `docs/architecture.md` now records that `/akinator` runs the complete pass by default, and the deliberate split between `akinator` (scales to the change) and `akinator-everything` (does not scale down). It had described neither. - `docs/deviations.md` item 1 now carries the amendment as well as the original six-to-one decision. ## [1.0.1] - 2026-08-26 ### Verified installing and loading in Claude Code Previously this was asserted by the structural tests and never actually done. It has now been done, against Claude Code 2.1.154: ``` claude plugin validate . -> Validation passed claude plugin marketplace add ./ -> added marketplace: akinator claude plugin install akinator@akinator -> installed (scope: user), enabled ``` `claude plugin details akinator` enumerates the whole surface: 21 skills, 7 agents, 1 command, 1 SessionStart hook, 0 MCP servers, 0 LSP servers. Always-on cost ~3.0k tokens - the per-skill always-on figure is the trigger description only, which is what makes 21 skills affordable; the body is paid on invoke. Two observations from the real install, neither a defect: - The details output reports "Skills (22)" and lists `akinator` twice. There are 21 skill directories on disk. The 22nd entry is `commands/akinator.md`, which shares the name - the inventory counts the command alongside the skills. - The install copies the whole repository into the plugin cache, so `.agents/`, `evals/`, `tests/` and `scripts/` ship with it. Only `skills/`, `agents/`, `commands/` and `hooks/` are loaded, and the generated `.agents/skills/` is **not** double-registered as skills. The rest costs disk, not context. ### Added - `docs/listing.md` gains the everything-pass use case - the flagship bare `/akinator`, which the listing had no example for. ### Changed - Version bumped across `.claude-plugin/plugin.json`, `.claude-plugin/marketplace.json` and `.codex-plugin/plugin.json`. The suite asserts all three agree. ## [1.0.0] - 2026-08-26 First release. ### Fixed before release - Codex plugin validation Three defects found by a validation run against the packaged plugin: - **`interface.composerIcon` and `interface.logo` are required**, not optional, and must reference square images that exist in the plugin. `docs/compatibility.md` had asserted the opposite and was corrected. Akinator now ships `assets/akinator-icon.png` and `assets/akinator-logo.png`, both 512x512, **generated** by `scripts/generate_assets.py` - the mark is drawn from a signed distance field and the PNG encoded with the standard library, so it keeps its provenance and can be drift-checked like every other generated artifact. - **Files directly under `skills/` are not imported** and fail validation. The skills index moved from `skills/README.md` to `docs/skills.md`, and `rules/08-skills-dir-holds-only-skill-directories.md` now prevents the class of defect with a test as its enforcement. - Both surprises are recorded in `memory/2026-08-26-codex-plugin-validation-surprises.md`, with the general lesson: **the validator is the contract; the field list is a summary.** ### The contract - The **creed** and the twelve-station **loop**, carried in the master `akinator` skill: ASK, RESOLVE, AUDIT, PLAN, IMPLEMENT, DOCUMENT, SKILLIFY, RULE, CONTEXTIFY, MEMOIZE, INDEX+SYNC, VERIFY. - Stations 6-11 are not deferrable, and the knowledge delta is declared **by path** at plan time - the structural choice that makes skipped documentation visible instead of silent. - The knowledge **taxonomy**: one canonical home per kind of knowledge, one home per fact, every artifact reachable from an index, generated beats written, and every doc states what would make it stale. ### Components - **`/akinator` runs everything.** With no arguments, or with free text, the command loads `akinator-everything` and runs the complete pass - every station, every applicable boardroom lens, every mechanical check, looping until the Definition of Done is proven with evidence. Mode words narrow the target, never the depth. See `docs/adr/0005-single-command-surface.md`. - **21 skills.** `akinator-everything` is the all-in-one pass; `akinator` remains the always-on router that scales the loop to the size of the change. The two are deliberately different settings - running the full pass on a typo is how a team learns to stop running any of it. - **20 station skills.** The loop's stations (`akinator-intake`, `-audit`, `-plan`, `-document-change`, `-skillify`, `-rule-forge`, `-contextify`, `-memoize`, `-adr`, `-index-sync`, `-router-sync`); business, product and operational mapping (`-business-map`, `-product-map`, `-ops-map`); discipline (`-gate-economy`, `-resource-guard`, `-anti-gaming`); installation and audit (`-onboard`, `-coverage`). - **7 boardroom agents** with real vetoes. `akinator-librarian` runs on every batch and blocks completion until stations 6-11 are satisfied. - **1 command**, `/akinator`, dispatching `onboard`, `audit`, `status`, `sync`, `question`, `decide`, and free-text work. - **1 SessionStart hook** injecting the behavioral contract and reporting which knowledge entry points the current repo actually has. - **10 templates**, each with a filled example written against one coherent fictional product so they cross-reference the way real artifacts do. ### Tooling - `scripts/akinator_coverage.py` - ten mechanically verifiable invariants: reachability, dead links, rule enforcement, router sync, module routers, generated artifacts, doc truth, skill format, staleness, and git hooks. Exit-code driven, JSON output, `.akinatorignore` support. **Never for a git hook**, and it says so in its own docstring. - `scripts/build_codex_pack.py` - deterministic generation of `.agents/skills/` and `AGENTS.md` from the canonical Claude skills, with a drift check. - `scripts/extract_components.py` - generates `context/components.md` from the tree; fails if a skill has no declared loop station. - `scripts/generate_assets.py` - draws and encodes the brand assets Codex validation requires, with a drift check. - `scripts/run_evals.py` - runs the behavioral evals against the fixtures. Fresh agent per step, no follow-up turn, each run in a disposable copy of the fixture so an eval can never mutate it, and optional grading by a second independent agent that sees only the transcript, the diff and the rubric. - `scripts/install-codex.sh` and `install-codex.ps1`. ### Verification - 130 structural and enforcement tests. Every rule's named mechanism is asserted to actually work, so a rule cannot silently become decoration. - Six behavioral eval suites, numbered without gaps, every one runnable by `scripts/run_evals.py`, against three fixture repositories: bare, brownfield with its own conventions, and one with deliberate rot. - CI runs the tests, all three drift checks, the coverage invariants against Akinator's own repository, a guard that the rotten fixture still fails, and a check that every eval suite is parseable and runnable. ### Behavioral eval results Four runs, recorded in `evals/results/`, including a **paired baseline** - the same task, same fixture, same model, with and without the pack installed. | Eval | Result | |---|---| | 01 silent-change (with Akinator) | **pass** - code, tests, and a README documenting the `billable` semantic that lived only in a docstring, plus a when-not-to and a stale-when line | | 01 silent-change (baseline, no pack) | **fail, expected** - code and tests; documentation explicitly declined | | 03 business-void | **pass** - refused to implement, filed seven dated open decisions in the fixture's own `docs/standards/`, asked the blocking question in business terms | | 04 newcomer | **pass** - identified the undocumented refund/quota void and refused to infer an answer from the implementation | The 01 pair is the product claim, measured: same model, one variable, and the knowledge artifact appears only with the contract installed. One paired run is a data point, not a study - the value is the trend across releases. Two defects were found by running them, neither visible from writing them: - **The installer shipped Akinator's own router into target repos**, giving them five dead links and an instruction to run Akinator's test suite. Fixed: the pack now generates a separate portable contract at `.agents/AGENTS.md` that names no repo-relative paths, and the installers copy that. - **The fixtures announced that they were fixtures**, naming their planted gaps in their own READMEs and contaminating any run against them. Fixed: those notes moved to `evals/fixtures/README.md` and each fixture now reads as an ordinary repository. Eval 04 should be re-run against the cleaned fixtures. Still unrun: 02 repeated-question, 05 gate-economy, 06 anti-gaming. The Claude Code plugin path - SessionStart hook, auto-triggered skills, boardroom subagents - was not exercised; the runs reached the agent through the Codex `AGENTS.md` delivery path and the plugin path remains verified structurally only. ### Platform contracts Verified 2026-08-26 against Claude Code 2.1.154 and the Codex skills and plugin manifest specs. See `docs/compatibility.md`. ### Deviations from the build brief Recorded in full in `docs/deviations.md`. The notable ones: - **One command instead of six**, at the owner's instruction (`docs/adr/0005-single-command-surface.md`). - **The Codex pack ships real Codex skills**, not loose prompt files - Codex has a first-class skills system reading `.agents/skills/`. - **No Codex hook surface**: Codex plugin validation rejects a `hooks` field, so the contract reaches Codex through the generated `AGENTS.md`. - **Gate receipts are specified, not shipped** - hook stacks differ too much between repositories for one implementation to be correct. Recorded as debt with a payoff condition.